Melanoscope AI decision support system for dermatoscopy

The purpose of the research is to present a medical decision support system using Melanoscope AI artificial intelligence methods for dermatoscopic diagnosis of skin neoplasms, implementing cascade and multiclass classification techniques, as well as a method for quantifying the interpretability of its models. Methods. The system includes a server subsystem for image analysis. The risk estimates for the routing algorithm are obtained based on the results of the ROC analysis of the first stage of the cascade. Results. On a subsample of 180 images, the greatest consistency with expert patterns was achieved through a mobile application for obtaining images and a visualization module for significance maps, which implements a cascading two-stage classification of dermatoscopic images. For four classification architectures (ViT-B/16, Swin, ConvNeXt, EfficientNet), significance maps were built using attention rollout methods for transformer and Grad-CAM for convolutional models. The consistency of the zones of high activation of the model with the expert annotation of clinically significant dermatoscopic structures, defined by the bounding box, is estimated by the intersection metric for integration (IoU) on a subsample of a clinically verified set of images. The thresholds for continuous architecture of ViT-B/16 (mean IoU 0.69 ± 0.12) versus 0.51 ± 0.15 for EfficientNet; the highest values were recorded for melanoma and basal cell carcinoma; Swin showed an intermediate result (0.64), convolutional architectures were lower (ConvNeXt 0.53). The differences between architectures are statistically significant (Friedman's criterion, p < 0.001). Thresholds P = 0.15 and P = 0.50 were obtained, defining a three-zone routing scheme for which clinical actions and the urgency of referral to a specialist in the clarifying nosological class of the second stage of the cascade were determined. Conclusion. The proposed method provides a reproducible quantitative assessment of the interpretability of cascade models and links continuous risk assessment with a specific clinical action. Transformer architectures provide higher consistency with expert patterns than convolutional ones, which justifies their choice for an interpreted dermatoscopic medical decision support system.